Claude Code 分层编排：Opus 顾问、Sonnet 主建、Haiku 蜂群

holy sh*t. Official Anthropic tip for Claude Code: stop burning Opus 5.5 context on tasks Haiku 5.5 can swarm, while Sonnet 5.5 builds put the architect on call with /advisor claude --advisor opus --subagents haiku Sonnet 5.5 drives the primary session on high effort, writing diffs and running test suites Haiku 5.5 subagents swarm the repo in parallel at 340 tok/s, handling file discovery and spec docs lookup Opus 5.5 stays on call in the background as the advisor, only stepping in at three critical moments: → before a plan locks: does the implementation plan miss auth invariants or schema contracts? → when a test breaks twice: are we fixing the root cause or falling into a recursive rabbit hole? → before calling done: did the full diff introduce hidden regressions or break pre-flight rules? Opus 5.5 reviews. Sonnet 5.5 builds. Haiku 5.5 swarms Jev engineering runs the identical pattern one layer down: mechanical decisions that need no reasoning (which file to open, which tool to invoke, retry or abort) execute in 16ms, so the frontier models only wake up when execution paths actually diverge Plan on high. Delegate on medium. Keep Opus on call. - the full advisor setup > Opus 5.5 on call reads full session history and catches deep architectural traps > Sonnet 5.5 lead drives edits, writes core logic, and executes test harnesses > Haiku 5.5 subagents swarm AST parsing, grep scans, and API docs in parallel > JEV micro-fork layer resolves 1,500+ mechanical routing branches in under 16ms > advisor stays completely silent on routine bash execution to protect context paste the setup and this prompt into Claude Code below: "Configure my Claude Code workspace for hierarchical advisor orchestration: 1. Audit ~/.claude/settings.json and project config for model roles fitting lead, subagents, and advisor: > Pin main session model: sonnet with effortLevel: high for primary execution > Pin subagents model: haiku with effortLevel: medium for parallel file discovery and docs retrieval > Pin advisor model: opus on call for strategic review 2. Enable the advisor tool in ~/.claude/settings.json: > Set advisorModel to claude-opus-5-5 > Enable subagent parallel dispatcher pool (3x workers) 3. Configure automatic advisor consultation checkpoints in ~/.claude/CLAUDE.md: > Consult /advisor opus before finalizing multi-file architecture plans > Automatically summon /advisor opus when the same test or compiler error fails twice > Enforce advisor diff contract audit before declaring tasks complete or staging git commits 4. Constrain subagent scopes: > Haiku subagents return structured AST summaries and docs snippets only without modifying main session context Show every configuration diff first. Do not apply edits until confirmed" ↳ https://t.co/gvAWk6KbDn

Harness Engineering: How to Build AI Workflows That Never Fall Apart Most people respond to a broken AI output by rewriting the prompt Then they switch models Then they upgrade to a larger context window The model still forgets instructions It still recommends the wrong tool It still skips verification It still claims success on broken work The problem sits outside the model weights The problem sits in the execution environment That environment is the harness Designing it is harness engineering Dario Amodei explained this directly during the launch of Claude Code: "Of course, you need an interface, you need a harness to use them" The model acts as a reasoning engine A model can predict the next logical step It cannot build a stable operating system on its own The harness controls what the model reads, what tools it can trigger, what data survives context resets, and when the execution must halt Prompt engineering adjusts the wording Harness engineering builds the infrastructure where that wording executes Put a frontier model in an empty chat box, and it generates casual prose Put that identical model in a repository with terminal access, automated test suites, project maps, and isolated workspaces, and it ships working software The weights stay identical The harness changes everything OpenAI recorded the identical pattern when training agents on Codex Their early runs failed because the environment was underspecified The engineering team avoided asking the model to try harder They built explicit capability checks and mechanical boundaries When an agent fails repeatedly, stop editing adjectives in the system prompt Inspect the operating system around the model --- # The Three Layers of a Reliable AI System Builders often treat prompt chains, agent loops, and multi-agent graphs as competing ideas They represent three distinct layers of one system ## 1. The Harness Layer (Environment and Memory) The harness consists of everything existing outside model weights It manages filesystem access, environment variables, authentication tokens, and durable files Andrej Karpathy established the defining mental model: The model is the CPU The context window is the RAM A raw CPU without storage or an operating system crashes instantly When users dump twenty pages of background text into a single prompt, they flood the RAM The model loses attention on critical rules A clean harness stores project knowledge in external markdown files It loads only the exact snippet needed for the immediate step ## 2. The Loop Layer (Evidence and Convergence) A model outputting code or text needs automated validation Asking the model "Are you sure this is right?" creates a circular hallucination loop The model evaluates its own probabilistic prediction and confirms its initial bias The loop layer introduces objective sensors: - Test runners (pytest, vitest, cargo test) - Format validators (JSON schema checkers, linters) - Execution budgets (maximum four attempts before human review) The model suggests the correction The harness runs the test and supplies the raw terminal output Execution continues only when mechanical evidence confirms the fix ## 3. The Graph Layer (Work Allocation) Long tasks crumble when forced into a single endless conversation thread The graph layer distributes work across isolated execution contexts One worker analyzes the task and writes an implementation blueprint A second worker writes the code in an isolated directory branch A third adversarial worker reviews the diff against project requirements Context stays clean No single prompt carries the burden of the entire project lifecycle --- # The 5 Core Jobs of a Personal AI Harness You do not need a massive enterprise platform to apply harness engineering A personal harness requires five mechanical rules in your everyday workflows ## 1. Turn Every Request into a Bounded Contract Conversational prompts allow models to silently redefine success When an ambiguous prompt encounters friction, the model solves an easier adjacent task and reports victory A production contract sets hard boundaries before generation starts: - Deliverable: Exact file name and format - Constraints: Dependencies allowed, style requirements, token budget - Non-Goals: Explicit list of items the model must avoid touching - Acceptance Criteria: The exact condition required to close the task Without an explicit contract, the model decides when it finishes With a contract, the criteria decides when the model finishes ## 2. Provide a Map, Not an Encyclopedia Context degradation strikes every frontier model over long horizons Dumping an entire project documentation file into chat burns tokens and dilutes focus Elite builders create a compact root map The map outlines where documentation, schemas, and source files live The model scans the map, selects the single relevant file, and reads that file alone A compact map preserves working memory A giant prompt dump exhausts it ## 3. Externalize Memory into Durable Files Chat history serves as temporary scratch space The conversation thread is not a database When a session runs long, attention fades and context windows fill up Store architecture decisions, user preferences, and unresolved bugs in a dedicated project file In Claude Code and Cursor, this file is `CLAUDE.md` or `.cursorrules` When starting a fresh session, the model reads the file and resumes immediately The work survives chat wipes, rate limits, and model switches ## 4. Build Sensors Before Granting Autonomy An agent cannot repair errors it cannot observe Telling a model "write bug-free code" changes nothing Providing a command that outputs exit code zero or a stack trace gives the model ground truth Sensors turn subjective quality into verifiable evidence: - Compilers and type checkers that fail on syntax errors - Automated test scripts that verify data transformations - Checklist validators that confirm all required fields exist The model creates the artifact The sensor generates the evidence The harness decides if the run passes ## 5. Apply the Dual-Encoding Rule Writing a rule in plain English leaves room for probabilistic failure The model reads the instruction, parses twenty other constraints, and skips the rule under load Dual-encoding solves this permanently: First, explain the rule in text so the model understands the goal Second, place a mechanical gate that blocks execution if the rule breaks If an agent must avoid deleting files, tell it in the prompt Then restrict its terminal permissions so `rm -rf` throws an operating system error The prompt guides the intention The mechanical gate protects the system --- # How to Set Up a Personal Micro-Harness Here is a five-minute harness structure you can paste into any repository, Claude Project, or Cursor workspace Save this content as `HARNESS.md` in your project root: ```markdown # Project Operating Contract ## Operating Boundaries - Allowed scope: Modify only files explicitly assigned in the current task - Forbidden actions: Never delete existing tests, never add unverified dependencies ## Project Map - /src: Application source code - /tests: Validation test suites - /docs/decisions.md: Durable log of accepted architectural choices ## Verification Protocol Before marking any task complete, run these checks in order: 1. Run local linter: `npm run lint` 2. Run test suite: `npm test` 3. Verify output format matches the requested schema ## Failure Escalation If a test fails twice on the same error: - Stop autonomous editing - Print the exact error log - State the proposed fix and wait for user confirmation ``` This short file replaces twenty lines of repetitive prompting in every chat It establishes project scope, controls file access, and forces mechanical verification --- # The Two-Pass Prompting Workflow When working on difficult reasoning tasks in standard chat interfaces, use a two-pass harness loop Most users ask for an answer and accept the first draft Production workflows split generation and critique into separate passes ## Pass 1: The Execution Pass ```markdown Draft the technical specification for [Feature X]. Follow the constraints listed in HARNESS.md. Do not evaluate your own output yet. Output the raw draft and state all assumptions made. ``` ## Pass 2: The Adversarial Review Pass ```markdown Review the draft above as a skeptical senior reviewer. Check against these specific failure points: 1. Are edge cases handled when input arrays are empty? 2. Does the logic introduce extra dependencies? 3. Does the solution violate any constraint from HARNESS.md? List every failure point found. Then output the revised version fixing those gaps. ``` Separating the creator role from the reviewer role cuts hallucinated claims in half The model cannot defend its own errors when explicitly instructed to attack them --- # Why the Harness Compounds Models update every few months Prompts that work today often produce different outputs on the next model checkpoint Harness engineering preserves your engineering effort When Anthropic releases an updated Claude or OpenAI launches a new reasoning model, your prompts might require minor adjustments Your test suites, project maps, permission boundaries, and durable state files remain intact They work on the new model immediately The model provides raw cognitive compute The harness provides the rails that turn compute into reliable execution Stop tweaking adjectives in your prompts Start building the harness around them